GPT-5.3 Codex

GPT-5.3 Codex

Quick answer: GPT-5.3 Codex is OpenAI's purpose-built agentic coding model — described as "the most capable agentic coding model to date" — optimised for use in the Codex environment. It offers a 400K-token context window, 128K max output, $1.75 input / $14.00 output per 1M tokens, and supports configurable reasoning effort. Knowledge cutoff is August 31, 2025. On LHTB it scored 0.203 mean reward (rank 20), solving 2 of 46 tasks.

At a Glance

Where GPT-5.3 Codex leads

• Specifically optimised for agentic coding in multi-step autonomous environments.
• Configurable reasoning effort (low / medium / high / xhigh) for precision-demanding tasks.
• Relatively affordable for an OpenAI coding model at $1.75/$14 per 1M tokens.

Where it lags

• LHTB score of 0.203 (rank 20) is well below newer models like GPT-5.5 and GPT-5.6.
• Knowledge cutoff of August 2025 is earlier than the latest GPT-5.x series.
• 400K context window is smaller than the 1M+ windows of later frontier models.

Best for: Autonomous coding agents in the OpenAI Codex environment and structured software engineering workflows.

What GPT-5.3 Codex Is

GPT-5.3 Codex is OpenAI's coding-specialised model in the GPT-5 series, positioned specifically for agentic coding tasks within the Codex product. OpenAI describes it as "the most capable agentic coding model to date" within its dedicated Codex prompting environment, where it can be guided through complex multi-step software engineering tasks using structured reasoning and tool-calling instructions.

The model supports four levels of reasoning effort — low, medium, high, and xhigh — allowing developers to dial in the trade-off between speed and depth of reasoning depending on the complexity of the task. Text and image inputs are supported, making it applicable to tasks that require reading UI screenshots or diagrams alongside code. Max output of 128K tokens allows it to generate large code artifacts in a single response.

Compared to its successors GPT-5.5 and GPT-5.6, GPT-5.3 Codex has a narrower context window (400K vs. 1M+) and an earlier knowledge cutoff (August 2025 vs. February 2026). On LHTB — which evaluates 18 models on 46 hard terminal tasks — it scored 0.203 mean reward, placing 20th of 21 models. This reflects both the model's earlier training and the fact that the benchmark's scoring harness (Terminus-2) is tuned for agentic performance that later models have improved.

Specifications

Related Benchmarks

• Long-Horizon Terminal-Bench (LHTB) — 0.203 mean reward, rank 20 (2/46 solved)
• SWE-bench Verified — standard coding agent benchmark